huzican0419/robotwin_piper_x_sim_real_from40k_step10000

Pi0.5 Piper sim/real — continued from 40k, checkpoint 10,000

BF16 PyTorch conversion of the OpenPI pi05_piper_sim_real_from40k checkpoint at global training step 10,000 from experiment robotwin_piper_x_sim_real_from40k.

Training

• Initialized from checkpoints/pi05_piper_new/robotwin_piper_x_new_sft/40000/params.
• Dataset: robotwin_piper_x_sim_real (8 tasks, 1,180 episodes).
• Reused normalization statistics from assets/pi05_piper_new/robotwin_piper_x_20_tasks_lerobot_v21_new.
• Cosine LR: warmup 500, peak 1e-5, decay 10,000 steps to 1e-6.
• Model: Pi0.5, action dimension 32, action horizon 50, discrete state input.
• Preprocessing: 14-dimensional state and actions, delta joint actions with absolute grippers, adapt_to_pi=False, task prompts and quantile normalization.
• Cameras: cam_high, cam_left_wrist, cam_right_wrist.

Export and inference

Converted using examples/convert_jax_model_to_pytorch.py with configuration pi05_piper_sim_real_from40k and BF16 precision.

• model.safetensors: inference weights.
• config.json: model configuration.
• assets/robotwin_piper_x_20_tasks_lerobot_v21_new/norm_stats.json: normalization statistics copied from the source checkpoint.
• conversion_info.json: source checkpoint and conversion details.

Use the OpenPI PyTorch implementation with the matching pi05_piper_sim_real_from40k configuration and the included normalization assets. This export does not include optimizer state for resuming JAX training.